Test-Time Training for Visual Foresight Vision-Language-Action Models
This work addresses the OOD vulnerability in visual foresight VLA models for robotics, offering a practical test-time adaptation method without architectural changes.
VF-VLA models are vulnerable to out-of-distribution shifts because action quality depends on predicted future images. The proposed T3VF method uses test-time training with an adaptive update filter to mitigate this, achieving robust performance with minimal extra inference cost.
Visual Foresight VLA (VF-VLA) has become a prominent architectural choice in the recent VLA due to its impressive performance. Nevertheless, the inherent design of VF-VLA makes it particularly vulnerable to out-of-distribution (OOD) shifts. Because the quality of action directly depends on the accuracy of the predicted future visual information, OOD conditions affect both stages at once. To address this vulnerability, we propose Test-Time Training Visual Foresight VLA ($T^3$VF), a test-time training approach motivated by the observation that the predicted future image and its subsequent observation form a natural supervision pair. To further address the practical challenges that arise from indiscriminate test-time updates, we introduce an adaptive update filtering mechanism. Empirically, $T^3$VF mitigates the OOD vulnerability of VF-VLA at a modest additional inference cost, without requiring any architectural modification or auxiliary modules.